Database and study population

We used data from the Dutch National Basic Registration of Hospital Care (LBZ).25 This database provides data from all 79 general and university hospitals in the Netherlands—at the time of the study period—and contains all hospital admissions. Dutch Hospital Data, the national organisation that administers the data from all the hospitals, gave permission to use the data anonymously. We selected index admissions with a discharge date from 1 January 2015 to 31 October 2016, and all subsequent readmissions until a discharge date of 31 December 2016. The data used in this study is fully anonymised and publicly available for researchers via Remote Access to Statistics Netherlands (CBS). We had permission of all hospitals to use the data anonymously.

The definition of a readmission was a clinical admission to the same hospital, within 30 days of discharge, following the clinical index admission—that is, the original hospital stay. We chose this time frame in accordance with the international literature.14 26 We calculated all-cause readmissions meaning that they do not need to be related to the cause of the initial hospitalisation.26 27 We used the index admission as the unit of analysis. This means that each readmission of the same patient is again an index admission for a subsequent readmission.28

Index admissions and readmissions were linked with a unique patient number obtained by a Trusted Third Party (Zorg TTP) which allows an individual’s information in healthcare to be exchanged without compromising their privacy. Readmissions were assigned to the hospital of the index admission. Transfers, which are defined as readmissions to another hospital within 1 day,29 were not counted as readmissions but included as an index admission of the second hospital.

We excluded hospitals that did not register unique patient numbers. We also excluded admissions that were not registered completely in the database (for example, missing diagnosis). Patients not living in the Netherlands were excluded as either their index admission or their readmission, could have taken place in their country of residence and therefore readmissions could be underestimated. Patients who died during their index admission were excluded from the population at risk. Furthermore, we excluded admissions where data were missing on one of the variables that we used in the analyses. Based on previous literature, we also excluded admissions in which the principal diagnosis involved either cancer care, obstetrics or psychiatric care.30

Hospitals with inadequate quality of data were also excluded. In order to assess the quality of data, we investigated the following criteria31: there should be at least 12 consecutive months of data registration; not >2% of vague diagnoses; at least 30% acute admissions; and at least 0.5 comorbidities, on average, per admission. We assessed these variables because they are subject to variations in coding between different hospitals31 and are important in the calculation of readmissions. Acute admissions and admissions with multiple comorbidities have a higher risk of readmission.1 13 Hospitals that did not meet one or more criteria were excluded from the analyses.

Design

We performed logistic regression analyses to calculate readmission ratios for each hospital based on the administrative data. We did not perform hierarchical modelling, as a recent study showed that adding a hospital level had only a very small impact on the results.32 The following predicting covariates for the adjustment for case mix were used33 34: severity of main diagnosis (a categorisation depending on the seriousness in terms of mortality), gender, age category, urgency of the admission, Charlson comorbidities (17 groups of comorbidity), socioeconomic status (based on the postal code of the patients’ residence), month of admission and place of residence before admission. All variables concern the index admission.

Patient and public involvement

Patients were not involved in the design of this study.

Analysis

We calculated the baseline characteristics of the subset of readmissions in the dataset, comparing these characteristics for readmissions to the same hospital with readmissions to other hospitals. We calculated readmission ratios for each hospital by dividing the observed number of readmissions by the expected number of readmissions, multiplied by 100. The expected number of readmissions is based on the case mix of the hospital. Two models were designed, one including only readmissions to the same hospital, while the other included readmissions to any hospital. We compared the readmission ratios of both models and calculated the correlation between both models with r.

We calculated 95% CIs for the readmission ratio of each hospital to analyse if it differed from the national average (readmission ratio of 100). Subsequently, we calculated the number of hospitals whose position of significance compared with the national average changed when taking into account readmissions to any hospital compared with the same hospital. A change in position of significance can be, for example, from significantly lower than the national average to no significant difference from the national average.

The models were calculated on the hospital level for all in-patients and in more detail on the level of medical specialties. The C-statistic of each model was calculated in order to determine the discriminative ability. We analysed the difference in C-statistic between the models including only readmissions to the same hospital, and the models with readmissions to any hospital, for each medical specialty.

Variables with fewer than 50 admissions in a category were merged with the smallest nearby category. This was done to prevent the SEs of the regression coefficients becoming too large. Comorbidities 9 and 17 (liver disease and severe liver disease), and 10 and 11 (diabetes and diabetes complications), were merged into one when there were fewer than 50 admissions where the comorbidity was present. Comorbidities with fewer than 50 admissions were not included in the regression analysis. We calculated the part of the readmissions to other hospitals for each medical specialty. Furthermore, we analysed which part of the readmissions to other hospitals concerned readmissions to general hospitals, leading hospitals undertaking clinical research and university hospitals.

The data were analysed using R V.3.2.3. The package pROC was used to calculate the C-statistic.

Comparison with other studies

We found 10.3% of admissions resulted in readmissions to any hospital, which is comparable with a study of Davies et al which came up with a figure of 10.1% all-cause readmissions.22 However, the Davies study was limited to acute care hospitals. In our analysis, we found fewer, 9.4%, readmissions when only looking at acute admissions and acute readmissions. Our analysis showed that 8.9% of the readmissions, both acute and non-acute, were in another hospital. This is low compared with the 17%–32% reported in other studies.16–23 These studies, however, concerned only acute care and were mainly carried out in the USA. When we limited our analysis to acute care, we found even fewer, 5.2%, readmissions to other hospitals. This might indicate that the impact of taking into account readmissions to other hospitals is not comparable across different countries with different healthcare systems.

For most medical specialties, we found C-statistics of the models with readmissions to the same hospital that were significantly higher. The largest significant difference was for cardiothoracic surgery. This indicates better prediction of the same hospital ratio compared with the any hospital ratio. However, Gonzalez et al 35 concluded that same hospital readmission rates provided unstable estimates of all-hospital readmission rates following coronary artery bypass grafting.

For some medical specialties, the C-statistics of the models with readmissions to any hospital we found were higher, with the largest significant difference for paediatrics. This indicates better prediction of the any hospital ratio compared with the same hospital ratio. A study by Kahn et al 36 also concluded that different-hospital readmissions differentially affect hospitals’ paediatric readmission rates. Our study found that 14% of the hospitals changed their position of significance compared with the national average when taking into account readmissions to any hospital compared with the same hospital. This is quite comparable with the finding of Kahn et al (2015) that excluding different-hospital readmissions incorrectly anticipated penalties for 11% of hospitals.

The Dutch healthcare system

The small amount of readmissions to another hospital might be caused by the strong gatekeeping and referral role played by GPs in the Netherlands. These GPs usually have consistent addresses for referring patients. Each hospital has a wide range of medical specialities, and each hospital delivers emergency as well as elective care. Some hospitals are specialised and deliver, for example, more complex care in the field of heart disease. However, when this concerns patients from other hospitals, it often concerns a transfer. Therefore, they are not taken into the analysis and do not have an effect on the readmission rate to any hospital.

The high level of patient satisfaction in the Netherlands can also be a reason for the low percentage of readmissions to another hospital. In contrast to patients in the USA, Canada, the UK or Switzerland, in the Netherlands, more patients report that their regular doctor has spent enough time on their consultation, has given explanations which are easy to understand and has involved them in decisions about care or treatment.37 This high level of patient satisfaction could result in Dutch patients usually going to the same hospital.

Strengths and limitations

We believe the current study is the first in the Netherlands that analyses the impact of taking readmissions to other hospitals into account. Our finding that the impact is much smaller compared with the literature, could also apply to other countries with a comparable healthcare system to the Netherlands.

Another strength is the completeness of the national administrative database which covers all hospital admissions. In this study, we used 2 333 173 admissions from 77 hospitals, which is 97% of the general and university hospitals.

A limitation of the study is that not all hospitals register the unique patient numbers completely. In some hospitals, a few per cent of the readmissions do not have a unique patient number. This affects the results from surrounding hospitals as when one of their patients is readmitted to another hospital that did not register the unique patient number, this readmission could not be taken into account. Therefore, the readmission rate of these hospitals could be underestimated. We decided not to exclude the hospitals with incomplete unique patient number registrations, because then the impact on the readmission rate of the surrounding hospitals would be much larger. However, we had to exclude one hospital from our analysis, because they did not register unique patient numbers for all admissions. We expect that this has a negligible impact on our overall findings; however, it does affect the results from the surrounding hospitals.

It should also be mentioned that the Dutch National Basic Registration of Hospital Care, the LBZ, does not contain a variable that distinguishes between intended and unintended readmissions. In the LBZ, we do have the variable ‘urgency’ (acute vs non-acute admission) that indicates whether care within 24 hours is needed.25 A recent study reviewed medical records of readmissions to evaluate the accuracy of a classification of potentially preventable readmissions with LBZ data.37 It appeared that a larger proportion of acute readmissions was classified as potentially preventable compared with non-acute readmissions (28.5% vs 5.0%). Nevertheless, we included both acute and non-acute admissions and readmissions in our study because complications might also result in readmissions that do not have a real 24-hour urgency and to avoid hospitals considering not to code the admission as acute in order to decrease their readmission ratio.

Implications for practice

Although the impact of taking into account readmissions to other hospitals is limited, this impact differs between hospitals. Therefore, these readmissions should be included in the readmission ratio, used in the Netherlands as a quality indicator, for a fair comparison between hospitals. However, its impact on the construct validity of the indicator is not known. It is important to include only readmissions that are related to the quality of care in the indicator and not readmissions that are a necessary part of the delivered care. Based on the results of this study, it is not certain if readmissions in other hospitals reflect substandard quality of care. Therefore, it is advisable to explore the readmissions in other hospitals by record reviewing to reveal the reason for readmission, before it can be decided if these readmissions should be part of the readmission indicator.

Besides, there are two concerns when applying this in practice.

First, hospitals cannot calculate their own readmission rate which includes readmissions to other hospitals. Therefore, a national organisation is needed that monitors the data from all hospitals in a specific country and which can apply case mix adjustment to readmission ratios, required if a fair comparison between hospitals is to be achieved.

Second, it is illegal in the Netherlands to share information about the readmission to another hospital with the hospital to which the patient was first admitted, without specific consent from the patient. This means that learning from readmissions to other hospitals is complicated.

As a result of these concerns, we advise not to take into account readmissions to other hospitals in the Dutch readmission indicator.

Future research

In order to identify areas for improvement, it is necessary to assess unintended readmissions. However, based on administrative data only, it is difficult to assess whether a readmission was unintended. Previous research showed that about 30% of the readmissions are potentially preventable.14 38 However, it is not known if this also applies to readmissions to other hospitals. Therefore, reviewing the records of readmissions to other hospitals is needed in order to analyse whether the readmission is a result of substandard care in the hospital where the original admission took place.

The group of patients who most often switch hospital, young men with relatively few comorbidities, may be interesting to explore further. For example, by using interviews to examine why they chose another hospital for their subsequent admission, in order to learn where quality can be improved.